DeepSeek open-sourced dsh-superagent on October 31, 2026, a prototype that turns GitHub requests into pull requests on your own machine. We compare it with Copilot's cloud agent, Claude Code's GitHub Action, OpenHands, Codex and SWE-agent.
DeepSeek's dsh-superagent, published under the MIT license on October 31, 2026, is a research prototype that listens to one GitHub repository, runs each accepted request in its own worker session and replies with a pull request. It is the only option here that runs entirely on your own machine with DeepSeek Harness. It is not production software, so GitHub's Copilot cloud agent or Claude Code's GitHub Action are safer choices for most teams today.
Key takeaways
The dsh-superagent repository was created on October 31, 2026 under the MIT license, per the GitHub API, and its README calls it a research prototype that is "not an official DeepSeek product".
Its design is a router session that decides whether a GitHub request becomes a Ticket, with each Ticket running in its own worker session and directory under a configurable concurrency limit.
DeepSeek's own docs warn that both deployment presets run with full access as the operator's OS user, so the agent should get a dedicated OS account and a separate GitHub user.
GitHub's Copilot cloud agent and the Claude Code GitHub Action run on GitHub-hosted infrastructure and bill through Actions minutes plus model usage, while dsh-superagent, OpenHands and SWE-agent are open-source tools you operate yourself.
SWE-agent's maintainers now recommend its successor mini-SWE-agent, so teams starting fresh with that lineage should look at mini-SWE-agent first.
Oct 31, 2026
dsh-superagent repository created (GitHub API)
MIT
License of dsh-superagent, OpenHands, SWE-agent and mini-SWE-agent (GitHub)
59 min
Maximum session runtime for Copilot cloud agent on GitHub Actions (GitHub Docs)
DeepSeek’s dsh-superagent, published on GitHub on October 31, 2026, is an open-source coding agent that watches one repository and answers requests with pull requests. It is a research prototype, and the hosted GitHub-native agents are the safer choice for most teams. The real comparison is between agents you rent and agents you run.
DeepSeek’s README describes a proactive agent for a single GitHub repository, packaged for DeepSeek Harness. This report compares it with five alternatives that take a GitHub request and return code, and says where each one fits.
Renting an agent versus running one
Copilot’s cloud agent and Claude Code’s GitHub Action run on infrastructure someone else operates or that GitHub’s runners provide. Both come with permission checks. Copilot respects branch protection and works on one branch per task. Claude’s action checks that the triggering user has write access and rejects bots unless allow-listed. dsh-superagent, OpenHands and SWE-agent put the agent on your own machine. That gives you control over code and models, but security becomes your job.
What makes dsh-superagent different
The Ticket model is the notable idea. A router session decides whether a request deserves a work item, and each accepted item gets its own worker session and directory. A web panel puts items needing a person first. The agent also starts proactive work in a record-only mode, so a team can see what it would have done before letting it act. These are design choices worth watching even if the prototype itself stays a prototype.
The security caveat
DeepSeek’s own deployment guide says both deployment presets run with full access as the operator’s OS user. The OpenHands README gives a similar warning for its no-sandbox mode. Giving an agent that can write code a dedicated OS account and a separate GitHub identity is the minimum, and review of every pull request should stay mandatory.
Which one to pick
Choose Copilot’s cloud agent if your team already pays for Copilot and wants the least setup. Choose Claude Code’s GitHub Action if you want to script triggers, skills and cloud routing. Choose OpenHands if you want an open-source, self-hosted setup that is not tied to one model vendor. Try dsh-superagent in a throwaway repository to see the Ticket design. Use Codex mainly for review and CI-fix loops on pull requests, and mini-SWE-agent if you want a small base to build on.
Ranked tools
01
1. GitHub Copilot cloud agent
Best for GitHub teams that want issue-to-PR with the least setup
GitHub's Copilot cloud agent works on delegated tasks in a cloud environment and can start from an issue, the agents tab, Copilot Chat or the GitHub CLI. GitHub's guide to working on an issue says some entry points open a pull request automatically. For others, you can ask Copilot to open one or create one from the session logs. Per GitHub Docs, it works on one branch at a time with one pull request per task. It honors repository and organization firewall allowlists in cloud sandboxes, and rulesets or branch protection can block it. GitHub Docs say Copilot in the cloud is available to paid Copilot plans starting October 28, 2026, subject to organization and enterprise policies.
Pros
Nothing to host, and built-in guardrails such as branch protection rules and read-only sessions without write access
Starts from where the work already lives, whether an issue, the agents tab or the CLI
Admins can disable it per repository
Cons
Works only with repositories hosted on GitHub
GitHub Actions sessions have a 59-minute maximum runtime and cover a single repository
Cloud sandboxes do not support self-hosted runners or custom compute
02
2. Claude Code GitHub Actions
Best for teams that want a scriptable agent inside their own CI
Claude Code GitHub Actions runs Claude Code inside your workflows, so mentioning @claude on an issue or pull request lets it analyze code, implement changes and push commits; the claude-code-action can also run on any GitHub event with a prompt. Anthropic's docs say runs consume GitHub Actions minutes plus API tokens, or your Claude subscription if you authenticate with an OAuth token. Only users with write access can trigger it by default, bots are rejected unless allow-listed, and enterprises can route inference through Amazon Bedrock, Google Cloud or Microsoft Foundry. Setup needs admin access to the repository.
Pros
Highly configurable, with skills, MCP servers, model choice and cron or event triggers
Write-access check and human-actor check run before every job
Cloud-provider routes for organizations that cannot use the direct API
Cons
You maintain the workflow files and secrets
Costs come from two meters, Actions minutes and model usage
Uses Claude models, so there is no cross-vendor choice
03
3. OpenHands
Best open-source, model-agnostic option you can self-host today
OpenHands is an MIT-licensed, open-source coding agent with roughly 92,000 GitHub stars as of November 2, 2026 (GitHub API). Its Agent Canvas control center is self-hosted, runs the OpenHands agent by default and can also drive Claude Code, Codex and Gemini CLI; its README says automations can integrate with GitHub, Slack and Linear, with an example of decomposing GitHub issues into tasks. It can run with a Docker sandbox, or with no sandbox at all, in which case the README warns the agent has full access to your filesystem. Setup needs Node.js 24 or later and uv.
Pros
Large, active open-source project with a pushed commit as recently as November 2, 2026
Not tied to one model vendor, and can orchestrate other agents
Docker sandbox option keeps the agent off your host filesystem
Cons
We could not confirm a turnkey issue-to-PR resolver from the pages we read, so GitHub automation may need configuration
Self-hosting means you own security, upgrades and model costs
No-sandbox mode gives the agent full filesystem access
04
4. DeepSeek dsh-superagent
Best to study or trial, not yet to rely on
dsh-superagent is a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. When an allowed person asks, a router session decides whether the request becomes a Ticket; each Ticket runs in its own worker session and working directory with a configurable concurrency limit, and the agent replies in the original conversation, for example with a pull request. A web panel lists items needing human attention first. It acts when addressed or when a team-written rule allows it, and proactive work starts in a mode that only records what it would have done. The repo's deployment guide lists Node.js 22.19+ or 24+, pnpm, git 2.32+, Python 3.8+ and tmux, SETUP.md adds GitHub CLI 2.48+, and GitHub Enterprise Server is not supported. The agent acts through its own GitHub user account.
Pros
MIT-licensed and runs on your own machine, with no third-party hosting of your code
Ticket-per-worker design with concurrency limits and a human-attention panel
Proactive behavior starts in a record-only mode, which limits early surprises
Cons
The README calls it "research code, not production-ready software" and says some designs in DeepSeek's article are not implemented
Both deployment presets run with full access as the operator's OS user, and the account wrappers are not a security boundary (per the repo's deployment guide)
It is tied to DeepSeek Harness, and the setup guide only says a model provider must be configured there
05
5. OpenAI Codex
Best for ChatGPT subscribers who mainly want PR review and fixes
Codex's GitHub integration centers on pull requests: connect a repository, turn on automatic review or comment @codex review and it posts a standard GitHub review that flags only P0 and P1 issues. Mentioning @codex with anything else, such as @codex fix the CI failures, starts a cloud task that uses the pull request as context and can push fixes to the branch if it has permission. OpenAI's page, as we read it, does not describe starting a pull request from a bare issue the way Copilot's agent does. OpenAI has also said that from December 2026 Plus, Pro and Business plans will draw ChatGPT and Codex from one shared token pool, per a report on OpenAI's help center update.
Pros
Strong fit for review and CI-fix loops on existing pull requests
Custom review rules can be set in AGENTS.md files
Bundled with ChatGPT plans
Cons
Less clearly an issue-to-PR tool than Copilot or Claude Code
OpenAI models only
Shared December allowance means heavy Codex use will compete with ChatGPT work
06
6. SWE-agent and mini-SWE-agent
Best for research and for building your own agent loop
SWE-agent is an MIT-licensed tool from Princeton and Stanford researchers that lets a language model of your choice use tools to fix issues in real GitHub repositories. Its README now says mini-SWE-agent has superseded it and recommends the smaller successor going forward, available from its own repository, which is also MIT-licensed and had a commit on October 26, 2026 (GitHub API). It is a command-line tool you run yourself, not a hosted bot that listens to a repository.
Pros
Model-agnostic and fully open source
Small, readable core that is easy to adapt
Long history in benchmark and research work
Cons
No GitHub-native listening, review panel or permission model
The original SWE-agent is superseded by its own maintainers' recommendation
You wire up triggers, sandboxes and PR creation yourself
DeepSeek's GitHub Agent vs 5 Alternatives (2026) — comparison
Tool
Runs where
How it starts
License
Maturity
GitHub Copilot cloud agent
GitHub cloud sandbox or Actions
Issue, agents tab, chat, CLI
Proprietary, paid Copilot plans
Product
Claude Code GitHub Actions
GitHub-hosted runners (or your cloud provider)
@claude mention or any GitHub event
Open-source action, Claude models
Product
OpenHands
Your machine or Docker sandbox
Agent Canvas UI and automations
MIT
Mature open-source project
DeepSeek dsh-superagent
Your machine, via DeepSeek Harness
GitHub request to an allowed agent
MIT
Research prototype
OpenAI Codex
OpenAI cloud
@codex review or @codex task on a PR
Proprietary, ChatGPT plans
Product
SWE-agent / mini-SWE-agent
Your machine, CLI
Command you run
MIT
Research tooling
Frequently asked questions
What is DeepSeek's dsh-superagent?
dsh-superagent is an MIT-licensed prototype coding agent that DeepSeek published on GitHub on October 31, 2026. It is packaged as a bundle for DeepSeek Harness, listens to one GitHub repository, and opens a work item for each accepted request, replying with a pull request.
Is dsh-superagent an official DeepSeek product?
No. Its README describes it as a research prototype and says it is not an official DeepSeek product. The README also says the published code differs from DeepSeek's internal version.
Is dsh-superagent safe to run on a work repository?
DeepSeek's deployment guide in the repository says both deployment presets run with full access as the operator's OS user and recommends a dedicated OS account and a separate GitHub user for the agent. Treat it as an experiment, not a production tool.
What is the best alternative for turning GitHub issues into pull requests?
For most teams on GitHub, Copilot's cloud agent is the simplest because it needs no hosting and can start from an issue and open a pull request. Claude Code's GitHub Action suits teams that want to script their own workflows.
Can I self-host a coding agent that works on GitHub?
Yes. OpenHands, dsh-superagent and SWE-agent or mini-SWE-agent are all MIT-licensed and run on infrastructure you control, but each needs more setup and your own security controls than the hosted options.
Is SWE-agent still maintained?
Its README says mini-SWE-agent has superseded it and recommends using mini-SWE-agent going forward.
How we evaluated
We confirmed dsh-superagent's creation date and license through the GitHub API on November 2, 2026, read its README, deployment guide and SETUP.md, and cross-checked a trade-press brief. For the alternatives we read each vendor's or project's own documentation or README on the same day. We did not install or run any of these tools, so we make no performance claims and cite no benchmark scores. Where we could not read a detail on the primary page, we say so. Rankings weigh readiness for real repositories, built-in guardrails and how clearly the tool turns an issue into a pull request, then openness and control.
Fit for the issue-to-pull-request workflow
Maturity and documented guardrails
Hosting model and control over your code
Model choice and cost transparency
Innvesti may have commercial relationships with some companies mentioned. Our analysis is
editorially independent. Figures cited are sourced; where reliable data is unavailable we
say so rather than estimate.
